The growing prevalence of prolonged isolation scenarios—from extreme environments to pandemic-related restrictions—has heightened the need for scalable systems to detect acute psychological distress. Traditional supervised approaches in mental health rely on labeled data, which are often scarce, biased, or slow to adapt. This paper presents a fully unsupervised machine learning framework based on the Iso-lation Forest algorithm to identify distress patterns using multidimensional behavioral metrics: perceived stress, mood variability, coping capacity, isolation duration, and sleep disruption. Using integrated public datasets, the iForest model achieved superior performance (Silhouette score = 0.61, Davies–Bouldin index = 0.89) and identified 6.3% of individuals as high-risk, a clinically plausible proportion. PCA visualization confirmed clear anomaly separation, and feature importance analysis highlighted confinement duration and coping difficulties as key predictors—aligning with established psychological theory. This work provides a methodological foundation for scalable, real-time mental health monitoring in resource-limited isolation settings. Future deployment requires clinical validation, multimodal data integration, and careful ethical governance for behavioral surveillance.
Johnson et al. (Sun,) studied this question.